Overall performance of Oort’s procedure for response shift detection at item level: a pilot simulation study

International audienceObjective. This simulation study was designed to provide data on the performance of Oort's procedure (OP) for response shift (RS) detection (regarding Type-I-Error, power and overall performance), according to sample characteristics, at item-level. A specific objective was to assess the impact of using different information criteria (IC), as alternatives to LRT (Likelihood-Ratio Test), for global assessment of RS occurrence. Methods. Responses to 5 binary items at two times of measurement were simulated. Thirty-six combinations of sample characteristics (sample size (n), "true change", correlations between the two latent variables and presence/absence of uniform recalibration RS (ur)) were considered. A thousand datasets were generated for each combination. RS detection was performed on each dataset following OP. Type-I-Error and power of the global assessment of RS occurrence, as well as overall performance of the OP was assessed. Results. The estimated Type-I-Error was close to 5% for the LRT and lower than 5% for the IC. The estimated power was higher for the LRT as compared to the AIC, which was the highest among the other IC. For the LRT, the estimated power for n = 100 and for the combination of n = 200 and ur = 1 item was below 80%. Otherwise, for other combinations of sample characteristics, the estimated power was above 90%. Conclusion. For the LRT, higher values of power were estimated compared to IC with appropriate values of Type-I-Error. These results were consistent with Oort's proposal to use the LRT as the criterion to assess global RS occurrence